mostly due to the increased size of the context window, and improvement to subagent parallelization.

But Claude will still spit out junk that requires refinement by someone who knows what they're doing. If Claude makes you obsolete in your job, you weren't doing anything hard.

These tools will certainly continue to improve, but it will continue to be incremental unless there's some kind of wild breakthrough on catastrophic forgetting or metallization, or some new idea enables more than just fancy interfaces with generative static models.